Abstract 2433: Uncovering molecular insights in breast cancer prognostics utilizing foundation model and digital pathology

Abstract Despite widespread use of clinicopathologic features and multigene assays, such as the Oncotype DX test, to guide clinical decisions, Breast Cancer (BC) remains a leading cause of mortality. There is a need to identify novel biomarkers to better prognosticate patients and ultimately influence treatment strategies to improve patient outcomes. In this study we developed an AI model for the prediction of BC distant disease recurrence (DDR) risk and applied it on the TCGA cohort to extract insights from gene expression data. Whole slide images (WSIs) of early HR+ HER2- BC cases were collected from Sourasky Medical Center. The median age was 61 years (range 25-75), median follow-up time 122.8 months (range 22.5-197.4) and median time to DDR 36.2 months (range 2.5-55.2). Utilizing a version of our foundation model, CanvOI, embeddings were extracted from WSIs, and a 5-fold cross-validation scheme was applied to develop the recurrence risk model. The model was then applied on the TCGA BC cohort WSIs, which included 98.8% females (n=1027), 55.3% HR+ HER2- (n=575) cases, and a median age of 59 years (range 26-90). Analysis of the WSIs of the TCGA BC cohort using the established model assigned a recurrence prediction score for each sample. The cohort was then divided into high-risk samples with the highest prediction scores and low-risk samples with the lowest scores followed by analysis of differentially expressed genes between the groups. The model development cohort included 291 H&E WSIs of female patients with early DDR (n=19, 6.5%) or no recurrence. 1039 TCGA WSIs were then analyzed with the established BC DDR risk model. Analysis of differentially expressed genes between the high AI risk group (n=100) and low AI risk group (n=124) was conducted. Notably, genes linked to estrogen receptor (ER) signaling pathways showed lower expression levels in the high-risk group, consistent with the established association of ER with more favorable outcomes, thus supporting the concept of this method. Further comparative analysis identified PTPRT as the second most differentially expressed gene between the groups, with lower expression in the high-risk group. Previous studies have associated PTPRT with tumor growth suppression, suggesting it may be an interesting candidate for further investigation. This proof-of-concept study highlights the potential of this approach of developing an AI model using a clinical cohort and applying it to an external cohort to extract insights from expression data. This enables the use of small clinical datasets with known outcomes to train models that can then be applied to larger datasets, such as the TCGA, which contain rich expression data, not accessible in the clinical data, to extract meaningful insights. This method effectively bridges the gap between limited clinical cohorts and expansive multimodal datasets, creating opportunities for broader and more precise discoveries. Citation Format: Shlomit Strulov Shachar, Tal Etan, 1 Didi Feldman, Ariella Deutsch-Lukatsky, Lee Ben-Itzhak, 1 Inbar Gilboa, Inbal Gazy, Ron Sinai, Addie Dvir, Reva Basho, Amir Sonnenblick, Dov Hershkovitz. Uncovering molecular insights in breast cancer prognostics utilizing foundation model and digital pathology [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2433.